Is US Chip Export Control to China Failing? The Rise of DeepSeek and a New AI Landscape

Is US Chip Export Control to China Failing? The Rise of DeepSeek and a New AI LandscapeOn February 12th, the US government attempted to curb China's AI development by tightening chip export controls, specifically restricting exports of Nvidia's most advanced products. However, the successful development of a generative AI application by Chinese AI company DeepSeek, comparable to top offerings from companies like OpenAI, powerfully challenges this strategy

Is US Chip Export Control to China Failing? The Rise of DeepSeek and a New AI Landscape

On February 12th, the US government attempted to curb China's AI development by tightening chip export controls, specifically restricting exports of Nvidia's most advanced products. However, the successful development of a generative AI application by Chinese AI company DeepSeek, comparable to top offerings from companies like OpenAI, powerfully challenges this strategy. It highlights the limitations of relying solely on restrictive measures to effectively counter the rapid pace of technological innovation. While DeepSeek's success, whose precise technical details remain undisclosed, doesn't entirely negate the role of US export controls in market and national security strategy, it clearly demonstrates that solely focusing on suppressing competitors cannot keep pace with technological advancements.

This event has sparked intense debate about the extent to which the US government should limit other countries' access to American chip technology. As President Biden's term neared its end, the Department of Commerce released new rules aimed at governing the global proliferation of AI chips and models. These new rules immediately drew sharp criticism from tech companies, including Nvidia, and numerous policy experts. Analysis from the Brookings Institution suggests these AI proliferation rules attempt to create a centrally planned global computing economy, raising concerns about their impracticality. John Villasenor, a non-resident senior fellow at Brookings and a professor of electrical engineering, law, public policy, and management at UCLA, bluntly stated: "Ten years from now, we'll look back and realize how unrealistic it was for the US government around 2025 to try to limit access to high-speed computing in 150 countries and territories.

Any future US countermeasures in the technological competition must confront a harsh reality: the idea of controlling innovation through export restrictions and similar methods is not guaranteed to succeed; it may even backfire. Brookings outlines potential risks, including stimulating a global AI ecosystem centered outside the US; encouraging more countries to strengthen technological cooperation with China; and even allowing non-US advanced chip manufacturers to gain larger global market share, weakening US companies' existing technological lead. Martin Chorzempa, a senior fellow at the Peterson Institute for International Economics, voiced concern: "I worry that we might overreact, over-regulate, without fully considering the costs and benefits.

Currently, a 120-day comment period for the AI proliferation rules exists, ending May 15th, unless President Trump rescinds or modifies the rule before then. While President Trump has generally emphasized the need to protect US technological leadership, he hasn't explicitly stated his support for this particular rule. The incoming administration's stance expanding, shrinking, or overturning existing chip export control rules remains unclear.

At the European AI Summit, Vice President JD Vance reiterated the US commitment to protecting its AI and chip technology and preventing its "weaponization." He emphasized: "Some governments steal and exploit AI to enhance their military intelligence and surveillance capabilities, acquire foreign data, and produce propaganda to undermine the national security of other countries. Let me be clear: this administration will stop that. We will protect Americas AI and chip technology from theft and misuse, and we will work with our allies and partners to strengthen and expand those protections, cutting off adversaries from access to AI technologies that threaten the security of us all.

President Trump's inaugural day executive order demanding to identify and eliminate vulnerabilities in existing export controls hints at a potentially hardline stance. The order mentioned that the government will "assess and recommend ways to maintain, acquire, and enhance U.S. technological superiority, and identify and eliminate vulnerabilities in existing export controlsparticularly those that permit the flow of strategic goods, software, services, and technologies to strategic adversaries and their proxies."

The US tech industry is actively engaging with the new administration. Executives from several major tech companies attended President Trump's inauguration, and Nvidia CEO Jensen Huang recently met with President Trump at the White House, discussing issues including chip export controls. Trump also stated that DeepSeek served as "a wake-up call to our industry, reminding us that we have to go all out to compete if we're going to win."

Of particular importance within the AI proliferation rules related to DeepSeek is the control of closed-source AI model weights, crucial in the training process for the AI systems thinking and response mechanisms. Chorzempa explains: "Part of the reason is that DeepSeek was able to circumvent the 2022 restrictions on chip sales to China, restrictions that were lifted in 2023 because the US realized they were wrong."

Chorzempa further points out that in 2022, when the US imposed chip export restrictions on China, the government set a specific parameter regarding the speed of inter-chip communication. The US reasoned that controlling only the performance of individual chips might not be sufficient because combining multiple lower-performance chips could still achieve near-supercomputer capabilities a level the US government didn't want China to reach. According to DeepSeek's description in its R1 model paper, the company seems to have successfully overcome this speed limitation. Chorzempa says, "At least by early 2023, experts in the field were already pointing out that as technology advanced, other restrictive measures would have to be put in place for effective control."

In 2023, the US government added further restrictions, making the Nvidia chips DeepSeek reportedly used for training no longer eligible for export. Subsequent moves by the Trump administration are likely to further tighten these restrictions. Yet, DeepSeek, working with a limited number of advanced chips and the innovations spurred by these limitations, succeeded in developing a better and possibly cheaper solution. Chorzempa notes: "DeepSeek seems to have cleverly optimized around the speed limits that were intended to constrain chip performance through software and hardware engineering."

AI Competition: Doing More with Less

The AI competitive landscape is evolving, with gaps narrowing in various aspects due to different factors. Alexandra Mousavizadeh, CEO of AI consultancy Evident, says: "The trend is actually that the gap between open-source and closed-source models is getting smaller. Today open-source models are becoming almost as powerful as closed-source models, while we see marginal costs of open-source models approaching zero."

DeepSeek has demonstrated that building viable large language models (LLMs) doesn't necessitate the most powerful computing power, with open-source technology serving as a viable alternative. Mousavizadeh argues that these restrictions can even act as a driver of innovation. She points out: These restrictions force researchers to use scientific methods, to compress data into smaller pools, and to employ techniques like Mixture of Experts (MoE) to achieve significant power reductions. Competitors in the AI field are able to do more with less."

Continuing such controls might accelerate China's domestic chip development. Mousavizadeh says: The dissemination of technology cannot really be stopped; the open-source environment is rife with technology sharing, and regardless of government policies, that sharing will happen.

If DeepSeek's success leads to further US tightening of advanced chip export controls, intended to slow Chinese AI progress, it should be clear that such measures are not a panacea. As Dario Amodei, CEO of generative AI startup Anthropic, recently wrote in a blog post: Export controls wont prevent competition between the US and China. Ultimately, if we want to win, US and other country AI companies will need to have more powerful models than Chinas companies, but we shouldnt give China a technological advantage unnecessarily. But Amodei expressed admiration for Chinese AI researchers: From their interviews, they seem to be bright, curious researchers whose real goal is to create useful technology.

Alex Karp, CEO of Palantir Technologies, in a recent interview stated: We have to run harder and faster, mobilize the whole society in pushing technology forward. Latecomers can catch up quickly, especially when we have done the innovation. Karp described DeepSeek as a derivative of US models and only marginally improved. He anticipates a massive policy discussion around technology exports to ensure US innovations dont hemorrhage abroad. But Karp added that ultimately, "the real advantage still belongs to the first-mover, as long as the first-mover keeps working... We are currently in the lead, and we must focus on ensuring that we remain in the lead. Palantir's rise wasn't by stopping others but by focusing on executing our own plans, which might be a thing to learn here."